The Agent Skills Directory

[DYNAMIC_EXECUTION]: The templates/monitor.py script uses importlib to dynamically load and execute Python code from a local file path (defaulting to .eval/default/graders/graders.py). This allows users to define custom evaluation logic (graders) but introduces a vector for arbitrary code execution if the local script is modified by an attacker.
[INDIRECT_PROMPT_INJECTION]: The skill's monitoring workflow ingests and processes external data which may contain adversarial instructions.
Ingestion points: The templates/monitor.py script reads production traces from .eval/default/traces/traces.jsonl (controlled by the --traces argument).
Boundary markers: The traces are expected to be in JSONL format, but the text content within those records is processed by graders without specific delimiters or safety warnings for embedded instructions.
Capability inventory: The loaded grader functions are executed on the trace data, which typically involves passing the content to LLM judges or other analysis tools.
Sanitization: There is no evidence of sanitization or filtering of the trace content before it is processed by the evaluation functions.

eval-maintain